MELP重要资料

-1-(Draft--May28,1998--Draft)SpecificationsfortheAnalogtoDigitalConversionofVoiceby2,400Bit/SecondMixedExcitationLinearPrediction1.INTRODUCTIONThisstandarddescribestheinteroperabilityrequirementsrelatingtotheconversionofanalogvoiceto2,400bits/sdigitizedvoicebyamethodknownasMixedExcitationLinearPrediction(MELP)andreconversionbacktoanalogvoice.Analgorithmdescriptionisalsoincludedtoaidimplementa-tionaswellasaperformanceverificationprocesstoverifyanimplementation.2.CONVENTIONSANDDEFINITIONS2.1FrameSizeAMELPframeintervalis22.5mspercentindurationandcontains180voicesamples(8,000samples/s).2.2AnalogSpecificationTherecommendedanalogrequirementsfortheMELPcoderareforanominalbandwidthrangingfrom100Hzto3800Hz.AlthoughtheMELPcoderwilloperatewithamorebandlimitedsignal,per-formancedegradationwillresult.ToensureproperoperationoftheMELPcoder,theA/Dconversionprocessshouldproducepeakvaluesof(ornear)-32768and32767.Additionally,thecodershouldhaveunitygain,whichmeansthattheoutputspeechlevelshouldmatchthatoftheinputspeech.3.ALGORITHMDESCRIPTION3.1CoderOverviewTheMixedExcitationLinearPredictioncoderisbasedonthetraditionalLinearPredictionCod-ing(LPC)parametricmodel,butalsoincludesfiveadditionalfeatures[1][2].Theseare:mixedexcita-tion,aperiodicpulses,adaptivespectralenhancement,pulsedispersion,andFouriermagnitudemodeling.ThesefeaturesareillustratedintheMELPdecoderblockdiagramshowninFigure1.Themixedexcitationisimplementedusingamulti-bandmixingmodel.Thismodelcansimulatefrequency-dependentvoicingstrengthusinganadaptivefilteringstructureimplementedwithafixedfilterbank.TheprimaryeffectofthismixedexcitationistoreducethebuzzusuallyassociatedwithLPCvocoders,especiallyinbroadbandacousticnoise.Whentheinputspeechisvoiced,theMELPcodercansynthesizeusingeitherperiodicoraperi-odicpulses.Aperiodicpulsesareusedmostoftenduringtransitionregionsbetweenvoicedandunvoicedsegmentsofthespeechsignal.Thisfeatureenablesthedecodertoreproduceerraticglottalpulseswithoutintroducingtonalsounds.Theadaptivespectralenhancementfilterisbasedonthepolesofthelinearpredictionsynthesisfilter.Itsuseenhancestheformantstructureofthesyntheticspeechandimprovesthematchbetweenthesyntheticandnaturalbandpasswaveforms.Italsogivesthesyntheticspeechamorenaturalquality.Pulsedispersionisimplementedusingafixedfilterbasedonaspectrally-flattenedtrianglepulse.Thisfilterspreadstheexcitationenergywithinapitchperiod,reducingsomeoftheharshqualityofthesyntheticspeech.0.01±-2-ThefirsttenFouriermagnitudesaredeterminedfromthepeaksoftheFouriertransformofthepredictionresidualsignal.Theinformationinthesecoefficientsimprovestheaccuracyofthespeechproductionmodelattheperceptually-importantlowerfrequencies.Thisincreasesthequalityofthesyntheticspeech,particularlyformalespeakersandwhenbackgroundnoiseispresent.3.2EncoderInputspeechisencodedbyperformingthefollowingstepsintheordergiven.3.2.1LowFrequencyRemoval.Thefirststepintheencodingprocessistoremoveanylowfre-quencyenergywhichmaybepresentintheinputsignal.Thisisaccomplishedwitha4thorderCheby-chevtypeIIhighpassfilter,havingacutofffrequencyof60Hzandastopbandrejectionof30dB.Thefilteroutputisreferredtoastheinputspeechsignalthroughoutthefollowingencoderdescription.Abuffercontainingthemostrecentsamplesoftheinputspeechsignalismaintainedintheencoder.Oneofthesesamplesisdesignatedthelastsampleinthecurrentframe.Thebufferextendsbeyondthissampleintothepastandfuturetocontainthesamplesneededfortheencodingprocess.Thelastsampleinthecurrentframeservesasareferencepointformanyoftheencodercalculations.3.2.2IntegerPitchCalculation.Forthispitchcalculation,theinputspeechsignalisfirstpro-cessedwitha1kHz,6thorderButterworthlowpassfilter.Theintegerpitchvalue,,isthevalueof,,forwhichthenormalizedautocorrelationfunction,,ismaximized.Thisfunctionisdefinedby:BandpassVoicingStrengthsFourierMagnitudesPitch&AperiodicInverseDFTShapingFilterShapingFilterNoiseGeneratorAdaptiveSpectralEnhancementscalePulseDispersionFilterLPCSynthesisFilterSynthesizedSpeechFigure1.MELPDecoderBlockDiagramFlagLSF’sGainP1ττ4041…160,,,=rτ()-3-,(1)where,(2)andrepresentstruncationtoanintegervalue.ThecenterofthepitchanalysiswindowisatsampleinEq.(2).Fortheintegerpitchcalculation,thiswindowiscenteredonthelastsampleinthecurrentframe.Thelowpassfilteroutputissamplewhenitsinputisthelastsampleinthecur-rentframe.Thetimeindexintheautocorrelationpreservesthepitchanalysiswindowalignmentarounditscenterpoint;thenormalizationcompensatesforchangingsignalamplitudes.Thefinalpitchcalculation(Section3.2.9)extendsthepitchrangetoalagof20samples.3.2.3BandpassVoicingAnalysis.Thisportionoftheencoderdeterminesthefivebandpassvoic-ingstrengths,,.Italsorefinestheintegerpitchmeasurementandthecorrespond-ingnormalizedautocorrelationvalue.Thebandpassvoicinganalysisbeginsbyfilteringtheinputspeechsignalintofivefrequencybands.Thesefiltersare6thorderButterworth,withpassbandsof0-500,500-1000,1000-2000,2000-3000,and3000-4000Hz.Arefinedpitchmeasurementismadeusingthe0-500Hzfilteroutputsignal.Thismeasurementiscenteredonthefilteroutputproducedwhenitsinputisthelastsampleinthecurrentframe.Twopitchc